SWE-Pruner is a tool designed to optimize the use of large language model (LLM) agents in software development, specifically addressing high token costs and latency issues. Traditional methods for context compression often overlook the specific needs of coding tasks. SWE-Pruner takes a different approach by allowing agents to set explicit goals and using a lightweight neural skimmer to identify and retain critical lines of code. This process involves formulating task-specific objectives and dynamically selecting relevant code lines, ensuring that essential details remain intact while reducing unnecessary information. The project boasts impressive performance metrics, achieving token reductions of 23-54% on the SWE-Bench benchmark and up to 14.84 times compression on LongCodeQA. It integrates seamlessly into multi-turn workflows, making it suitable for complex software engineering tasks. The model, which has a size of 0.6 billion parameters, highlights semantically important lines of code while preserving logical structures. This adaptability makes it useful for various scenarios, from debugging to feature development. The repository includes detailed directories for experiments, evaluation benchmarks, and code utilities. Users can find an inference tutorial to help them get started, while developers can access training scripts to create their own pruners. Researchers have scripts available for reproducing results and are encouraged to use a multi-GPU setup for efficiency. The project has been supported by teams from Bytedance and Alibaba, and relevant papers and code are linked for further exploration.